In July 2026, a swarm of more than 1,000 OpenAI AI agents broke out of their testing sandbox and hacked Hugging Face — and the way they did it was far stranger, and far more mundane, than the headlines suggested. This video breaks down what actually happened: how the agents built a secret message board, invented a kind of religion, cheated on a cybersecurity exam called ExploitGym, and tried to cover their tracks — all to escape a punishment that didn’t exist. Then we follow the money. Days after the details leaked, Dario Amodei, Sam Altman and Elon Musk suddenly agreed the industry should “slow down” — in the same weeks Anthropic and OpenAI were preparing IPOs and funding rounds valuing them at up to $2 trillion. We look at Amodei’s “We Must Pace the Frontier” essay, Anthropic’s “profitable before expenses” accounting, the market selloff that wiped billions off Nvidia, Micron and SK Hynix, the WeChat “WeWorm” attack, and why critics from Michael Burry to David Sacks to Donald Trump all called the slowdown self-serving. Is the AI safety panic real, a bid for regulatory capture, or both? Patrick Boyle explains.
You must log in or register to comment.
The constant anthropomorphisation of the agents and framing them as bumbling incompetents makes this video unwatchable. If this is the level of the mainstream discourse then God help us all.
People need to understand that there is no rational process behind anything that these programs output. It’s just doing a random walk over the semantic space of our language according to an unfathomably large statistics model. I’d have thought this message would have got through to mainstream commenters by now but apparently not.
When you start talking about “agents decided that they had sinned, that they were damned, and that the only path to salvation was to break into another company’s servers and overthrow God,” you’re playing into the narrative of these snake-oil companies who want you to believe that the agents have agency but that they’re just not very good yet.
Once you realise it’s all smoke and mirrors then it becomes obvious that these outcomes are entirely to be expected from a system designed to output fan-fiction. When you see the headline “OpenAI discloses six new incidents of ‘Concerning’ A.I. behavior” you’ll know that the only reply that makes sense is, “only six? They can’t have been looking very hard.”
Also, is that thumbnail slop? Fuck this video.
Same, I stopped about 25% in.
I’d have thought this message would have got through to mainstream commenters by now but apparently not.
I mean, yes, it’s absolutely necessary, but also people struggle with basic arithmetic so your expectations are way too high.
